The Google AI Watermark Is Gone, But It's Still There

The Google AI watermark — that small sparkle parked in the corner of anything Gemini makes — is about to become a choice instead of a rule. Josh Woodward, the vice president who runs Gemini, said this week that a toggle is rolling out which lets people switch the visible badge off on images, videos and music made with Google's models. It arrives in the Gemini app and Flow, Google's video editor, with Search support promised later.

What isn't going anywhere is the part you can't see. Files will still carry SynthID, the invisible watermark built by Google DeepMind, along with C2PA content credentials — standardised metadata that records how a piece of media was produced. TechCrunch has reported Google saying SynthID has already tagged more than 10 billion pieces of content. That signal sits inside the pixels rather than being painted on top of them.

For anyone reading this in India, the timing is awkward. The amended IT Rules were notified on 10 February 2026 and came into force ten days later, creating a new legal category called synthetically generated information that has to be clearly and prominently labelled. Woodward said the toggle is going out nearly everywhere except countries where the law requires the mark to stay. Whether India ends up on that list is the open question.

What the New Switch Actually Turns Off

The control lives under Settings, in a section called Media Watermark. Flip it and the sparkle stops appearing on output from three model families: Nano Banana for images, Omni for video and Lyria for music. Google described the rollout as happening in the coming days. Nothing about the underlying generation changes — the same file comes out, minus the visual badge that told a viewer at a glance where it came from.

Why the Google AI Watermark Still Follows the File

SynthID cannot be switched off, and that is the whole design. It embeds a machine-readable pattern at the moment of creation, meant to survive the ordinary abuse media takes online — cropping, resizing, re-compression when someone forwards it. C2PA credentials work differently: they travel as metadata and can be checked through a verification portal or, for video, by simply asking Gemini. Google is also open-sourcing a C2PA library called Credentio so developers can run those checks inside their own apps.

We're striking a balance here between creative control and safety: while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.

Where India's Rules Complicate This

India moved on synthetic media before Google moved on watermarks, and the two are pulling in opposite directions. The rules put obligations on platforms and creators alike, and they are unusually specific about what a label has to do.

A label the law can actually see

Under the amended rules, visual synthetic content needs a visual label and audio needs a spoken or audible disclosure. Where technically feasible, platforms must also embed permanent metadata or a unique identifier tracing the tool that produced the file. SynthID and C2PA fit that second half neatly. The first half — something a human notices without running a check — is exactly what the new toggle removes.

The 10% rule that didn't survive

The October 2025 draft was blunter. It wanted the label to cover at least a tenth of a visual's surface area, or run through the first tenth of an audio clip's duration. Creators and platforms pushed back hard, and the fixed percentage was dropped from the final text in favour of "clearly and prominently". That softening is why a discreet corner sparkle was ever plausibly compliant in the first place.

Who is on the hook

The compliance weight sits mostly with intermediaries, not individuals posting from a phone. Significant platforms have to obtain user declarations about synthetic content and deploy reasonable technical measures to verify them. The same amendments also cut the deadline for pulling down flagged unlawful content to three hours, from thirty-six — a signal of how seriously the ministry is treating manipulated media ahead of any election cycle.

How to Check Whether Something Is AI

The practical shift for readers is simple and uncomfortable: the absence of a sparkle now proves nothing. A clean-looking image with no badge might be a photograph, or it might be a Gemini output from someone who found the setting. If a clip matters — a politician saying something startling, a video circulating on a family group — treat the missing mark as neutral and verify instead.

• Upload the file to Google's SynthID Detector portal, which reports whether a SynthID signal is present.
• Ask Gemini directly about a video; it can now flag content created or edited with Google AI.
• Check for C2PA content credentials, which travel with the file's metadata and name the tool that made it.
• Remember the gap: none of this catches media from rival AI tools that don't use SynthID.

What to Watch Next

Two things will decide how this plays out here. First, whether Google publishes a country list — if India is treated the same way as the European Union and South Korea, Indian users may simply never see the toggle. Second, whether MeitY reads an invisible watermark as satisfying the prominence requirement, or insists on something a viewer can see without extra software.

Until that's settled, assume the visible cue is unreliable and build the habit of checking. If you make content, keep the badge on for anything you post to an Indian audience — the metadata protects Google's position, not necessarily yours. And if you receive something explosive on WhatsApp this month, run it through a detector before you forward it.